The current focus of the world\u2019s governments, business and media when it comes to artificial intelligence (AI) is on the dangers it poses to jobs, democracy and even the future of humanity itself. For many of a certain generation, these fears crystallise in the cold but relentlessly courteous voice of HAL, the on?board AI in Stanley Kubrick\u2019s 2001: A Space Odyssey. After a crew member questions HAL\u2019s diagnosis of a faulty antenna, the system secretly lip?reads the astronauts, discovers their plan to disconnect it, and kills them one by one. Nobody who sees the film is likely to forget the chilling reply from HAL to last surviving crewman Dave Bowman\u2019s order to \u201copen the pod bay doors\u201d so he can re-enter the spacecraft after being locked outside by HAL during a spacewalk: \u201cI\u2019m sorry, Dave. I\u2019m afraid I can\u2019t do that.\u201d As it stands, though, potentially far more chilling for U.S. AI firms is the huge structural disadvantage that they have compared to their Chinese counterparts and the massive drop in their companies\u2019 stock market valuations that this may soon catalyse.<\/p>\n
\u201cThe key problem here is that the major Chinese AI players currently achieve around 90% of the performance of their U.S. competitors but at only about 10% of the cost,\u201d Mehrdad Emadi, head of risk analysis and energy derivatives markets consultancy Betamatrix, in London, exclusively told OilPrice.com last week. \u201cUp to half of U.S. AI firms\u2019 costs is the electricity needed to drive the data centres, and this is only likely to increase from here,\u201d he said. \u201cBut the much lower costs of their Chinese AI rivals are likely to stay where they are, and may even edge lower,\u201d he underlined. \u201cMeanwhile, the performance of China\u2019s AI is likely to keep increasing to a point where it nears 100% of its U.S. rivals,\u201d he added, \u201cand that looks like a deadly convergence for American AI firms.\u201d Underpinning this huge discrepancy in costs between U.S. and Chinese AI firms is the fundamental difference in the way the electricity power grids were designed and built in the first place, highlights Steve Keen, honorary professor at University College London and the economist who most cogently warned that the economic crisis that began in 2007 was imminent. \u201cChina benefits from its centralised power grid that was designed from the outset by engineers rather than the decentralised one in America designed by accountants and similar,\u201d he exclusively told OilPrice.com. \u201cThe American grid may have been cheaper to install initially, but it can\u2019t transmit energy as widely across the country as the Chinese system, and it loses much more power than the Chinese grid when it does so,\u201d he underlined. Related: Saudi Arabia Restarts East-West Oil Pipeline<\/a><\/p>\n More specifically, China\u2019s grid operates under a unified national strategy controlled by the State Grid Corporation, allowing power to move seamlessly across thousands of kilometres. By comparison, the U.S. grid is split into three isolated interconnections — Eastern, Western, and Texas\/ERCOT — which makes moving power far more difficult and costly. To increase the amount of energy transmitted at any given time, operators must either increase current or voltage. In this context, China\u2019s power grid is built on an extensive backbone of 800-1,100 kilovolt (kV) Ultra-High Voltage lines utilising Direct Current. This combination enables it to move up to 12 gigawatts (GW) of power — the output of 10 nuclear power plants — down a single transmission corridor over 3,000 kilometres with almost zero power loss. By contrast, the U.S. grid can get nowhere near this level of total power generation, or near-zero power loss over distance, as it relies on standard 345-500 kV trunk lines in just a High-Voltage Alternating Current set up. In sum, the U.S.\u2019s fragmented, aging grid acts as a physical ceiling on how large and how fast American tech companies can scale their AI clusters, whereas none of these constraints apply to China\u2019s system — or, by extension, to its AI firms.<\/p>\n It has been widely conjectured that these surging power demands in the U.S. could be met by increased oil and gas flows, or supply from other sources. However, on the renewable energy side, solar and wind power still fail to provide the \u2018always-on\u2019 supply that AI data centres demand, given the intermittency of both sources. Nuclear power — including small modular reactors (SMRs) — has become a recent focus, with U.S. hyperscalers including Microsoft, Amazon, and Google eyeing tens of billions of dollars in investment in such projects. Critically, though, there is a massive time lag and potentially huge additional costs attached to both mainstream and SMR nuclear projects. Traditional nuclear projects in the U.S. have routinely suffered from multi-billion-dollar cost overruns and 15-year development timelines. SMRs, despite lower upfront costs, still target construction windows of up to seven years. Moreover, the need for regulatory approvals from agencies such as the U.S. Nuclear Regulatory Commission, combined with physical construction constraints, means the commercial SMR sector is unlikely to hit scale until the late 2030s. On the fossil fuel side, oil remains too valuable to burn for baseload grid power, leaving natural gas as the obvious main feedstock for the AI power grid going forward. \u201cBut the problem here,\u201d said Emadi, \u201cis that everyone\u2019s after the same equipment at the same time to convert it into electricity for the grid, and there are very few suppliers who make what\u2019s needed.\u201d A prime example of this, he highlighted, is the gas turbine manufacturing market, which is dominated by the \u2018Big Three\u2019 oligopoly of the U.S.\u2019s GE Vernova, Germany\u2019s Siemens, and Japan\u2019s Mitsubishi Heavy Industries. \u201cOrder backlogs are sold out, waiting times are up to seven years and growing, and prices are on track to triple,\u201d he underlined.<\/p>\n None of these structural hindrances are likely to improve any time soon, with colossal increases in U.S. AI power demand forecast in the coming decade. According to Goldman Sachs Research, U.S. data centre capacity demand is forecast to climb from the current baseline of around 42 GW to a range of 118-134 GW by 2030, while long?term modelling from BloombergNEF projects a rise to around 194 GW by 2035. \u201cWith American companies producing high-quality AI, but at a much higher cost than Chinese companies which produce products of very comparable quality, the US firms have to start generating revenues to match, and to do this they need to boost their steady power supply enormously,\u201d said Keen. \u201cBut this looks extremely difficult to do given the physical power constraints, so the potential for U.S. AI firms\u2019 valuations to fall looks significant,\u201d he added. \u201cAnd what\u2019s happening now in U.S. AI is typical of an industrial bubble in a capitalist economy, which always causes booms and busts, as happened for example in the U.S. railways bubble from the 1870s, which was high-tech at the time,\u201d he told OilPrice.com. \u201cEarly-mover railway firms were flooded with interest and huge overinvestment occurred, but ultimately only a handful of those railways made profits, with most ending up losing money and being sold for a song in the slump that followed the boom,\u201d he underlined.<\/p>\n It could be even worse in the case of a U.S. AI bubble, Emadi told OilPrice.com, because of the heavy involvement of \u2018private credit\u2019. \u201cTo get around the rules on riskier lending by banks that were put in place after the 2007\/08 financial crisis, several big AI firms used their own huge cash reserves to invest in their AI projects, but they also borrowed money from non-bank lenders — the private-credit sector — to ramp up that leverage,\u201d he said. \u201cSo we\u2019re talking about hundreds of billions of debt attached to these American AI outfits, which have costs running at around five to ten times their revenues,\u201d he added. \u201cWhen the AI firms start to go bust, these private-credit lenders will have to sell other assets to try to compensate for their losses, and, given the size of the exposure, this is likely to start a cascade effect of selling, lower prices and more selling through the financial system, crashing valuations and catalysing bankruptcies,\u201d he added. \u201cThis won\u2019t be a market readjustment any more than being shot in the head is a physical readjustment, with my guess being that we\u2019re going to see 35-50% of the capital valuation of the entire US AI sector disappear very fast,\u201d he added. \u201cThat includes all but two or three of the big players going bankrupt, as they may get bailed out by the U.S. government because of their close involvement in the defence sector — otherwise, I see most of the other AI companies that survive being left with a value of around 11 cents to the dollar,\u201d he concluded.<\/p>\n By Simon Watkins for Oilprice.com<\/p>\n